Marine debris detection in a coastal vegetated area using UAVs

Plastic has become one of the most ubiquitous materials produced by humans in the past century. Society has discovered many uses for it, but despite recycling programs, much of it ends up discarded in the natural ecosystem. Plastic pollution is having negative impact on our oceans and wildlife, causing accidental intake or by mistaking plastics for preys, leading to intestinal occlusion, suffocation or immobilization. For instance, the toxins from the plastics have entered the food chain, threatening human health. There are many challenges with plastic pollution in both coastal and marine environments mainly due to the size of the plastic pieces. In this study, we aim to aid in the understanding of how marine plastics move in coastal areas and help facilitate cleanup efforts by mapping the trash in a variety of environments affected by plastic waste. We use recently developed machine learning algorithms to detect the presence and location of litter in images captured by a multi-spectral camera mounted on a UAV. There are many challenges associated with even collecting data on plastic pollution in both coastal and marine environments. The main objective of this project is to develop an algorithm that can be loaded onto a drone and detect trash in real time; we implement a simple and efficient workflow to process the image data. After the data has been processed, we feed it into a neural network. Two neural network algorithms are tested to see which works the best for this application. The "Faster R-CNN" software uses 2 neural nets: one to find bounding boxes of objects in the images, and another to classify each object that it finds. The SSD algorithm uses a single net to both detect and classify objects, searching through a predefined set of bounding boxes and adjusting them. However, since the multi-spectral camera we used can only capture one image per second, we think that the increased accuracy of the Faster R-CNN algorithm is ideal for this application. The classifier trained with images from different environments and a wide a variety of objects with different shapes, reflections and colors. This allowed us to make the neural network more robust and less prone to over-fitting. The data was tagged by hand, but once the bounding boxes are detected more accurately, additional tagging is possible with significantly less human effort.